Reproduce “Vid-SME: Membership Inference Attacks against Large Video Understanding Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qi Li, Runpeng Yu, Xinchao Wang
Vid-SME introduces a grey-box membership-inference attack tailored to video-language models. It compares Sharma–Mittal entropy from videos shown in forward and reversed temporal order, while adapting entropy parameters to motion and illumination; on several VULLMs, the method improves detection, reaching AUC 0.840 and TPR@5% FPR 0.420 on Video-XL-CinePile-7B with 16 frames.
Key insights
- Training-member videos tend to show a larger loss of predictive concentration when their frame order is reversed, providing a temporal signal that static membership metrics miss.
- Adaptive entropy parameters are useful because high motion and changing illumination alter the sharpness and reliability of token distributions differently across videos.
- Attack success is not determined solely by the model’s train–test performance gap; the number of sampled frames can change how distinguishable member and non-member behavior becomes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.